The other full-stack AI company
AMD is no longer the PC-chip underdog of the 2010s. It is a roughly $890 billion public company whose Data Center segment just doubled year over year and now brings in most of the revenue. The bet is that the AI buildout wants a second source besides Nvidia: an open rack, an open software stack, and a CPU franchise that already runs a large share of the world's servers.
That is a different shape from Nvidia. AMD still sells Ryzen PCs, Radeon GPUs, console SoCs, and Xilinx FPGAs. But the story that moved the stock from the teens to $545 is Instinct plus Helios plus multi-gigawatt contracts with OpenAI, Meta and Anthropic. Whether those contracts turn into 2026 and 2027 revenue is the only question that currently matters.
From a second source to Zen
AMD was founded in 1969 as a Fairchild spin-out and spent its first decades as Intel's licensed x86 second source, then as a clone-maker after the 386 lawsuit. Athlon and Opteron were real hits in the early 2000s; the late 2000s were not. The ATI acquisition in 2006 brought graphics, the GlobalFoundries spin-off in 2009 made AMD fabless, and a near-death stretch in the early 2010s was cushioned by PlayStation and Xbox SoCs.
Lisa Su took over in 2014. Zen and Ryzen in 2017, then EPYC in the server hall, rebuilt the company on process from TSMC rather than a captive fab. Xilinx in 2022 added FPGAs. Instinct, once an also-ran HPC GPU, is now the product hyperscalers will pay gigawatts for if the software holds.
Helios is the product, not just the GPU
Nvidia sells an AI factory. AMD's answer is Helios: 72 MI455X GPUs, 18 Venice CPUs, Pensando networking and ROCm in an Open Rack Wide cabinet co-designed with Meta's OCP work. The claimed edges are memory (432 GB HBM4 per GPU, 50% more than Vera Rubin at rack level) and cost per token, not CUDA compatibility.
The software gap is the historic reason this pitch failed. AMD is trying to close it by putting OpenAI, Meta and Anthropic inside the compiler loop, launching ROCm.ai so coding agents speak ROCm, and pointing at MLPerf gains on the same MI355X hardware six months apart. Three million models running out of the box is the figure Lisa Su cited on the Q2 call.
Contracts measured in gigawatts
OpenAI's 6 GW deal, Meta's matching 6 GW including a custom MI450, and Anthropic's 2 GW are how AMD is buying a seat at the frontier-lab table. The OpenAI and Meta warrants (160 million shares each at a penny, vesting on purchases and a $600 stock price) align those customers with AMD's equity. Anthropic instead takes up to $5 billion of AMD capital and gives AMD Claude for ROCm work.
Oracle's 50,000-GPU public cluster is the most concrete near-term unit count. Microsoft is putting Helios on Azure for inference. HUMAIN already turned on MI355X in Saudi Arabia. That roster is still a fraction of Nvidia's, and several of the gigawatts do not start until 2027.
What to watch
The next 90 days decide whether Helios is a slide or a shipment. Q3 guidance of $13 billion implies another sequential jump; if Helios misses late Q3, that print and the Q4 ramp become the slip. OpenAI's first gigawatt, Meta's custom GPU, and Oracle's 50,000 GPUs are the named proofs.
Further out, MI500 in 2027 has to land while Nvidia is already on Rubin and custom silicon is taking inference share. AMD also has to keep EPYC growing as the agentic-CPU story, because the company is still a CPU vendor that happens to be winning GPU sockets, not the other way around. China remains a capped Instinct market after the MI308 export-control charges in 2025. The bull case is that the AI factory market is large enough for a strong second place with better memory and a lower bill. The bear case is that CUDA plus an annual Nvidia cadence plus customer ASICs leave AMD a rounding error at 5-7% of accelerator revenue even after Helios ships.